HostoSink: A Collaborative Scheduling in Heterogeneous Environment

摘要

Due to the limitations of power consumption and memory capacity, the past few years have observed a strong trend of using heterogeneous environment equipped with accelerators, such as GPU (Graphic Processing Unit) and FPGA (Field Programmable Gate Array), and even MIC (Many Integrated Core), to help the traditional SMP (Symmetric Multi-Processing) CPU to speed up applications. In this paper, we choose the Intel MIC architecture coprocessor as the accelerator and design HostoSink, a runtime system for collaborative scheduling based on Pthread task. With the help of runtime characteristics of the application and the heterogeneous environment for scheduling the Pthread tasks between CPU and MIC automatically and dynamically, HostoSink provides MIC users with an easier way to gain high performance in heterogeneous CPU-MIC environment without the need of optimizing the original Pthread-based multi-threaded applications manually too much. Experimental results show that by using HostoSink, the overall speedup can achieve more than 3x speedup compared with the original performance by using CPU only and the average amount of data transmission between CPU and MIC is also reduced.